Experimental beta · local-first edge monitoring
One observation becomes local evidence.
Sentinel watches one camera from an NVIDIA Jetson. It detects and tracks people, applies zone rules, and keeps an incident record on the device beside the camera, where you can inspect it.
00.0 sWaiting for the next arrival
Synthetic sample · a scripted loop drawn in your browser · no camera connected
01 — What it is
A frame, a rule, a record.
A camera captures what happened. Sentinel connects the observation, the rule and the evidence on the hardware beside it.
-
01
Start with the source.
One RTSP stream enters the local pipeline. Every observation carries its frame's source time, the boot and the stream epoch, so a stale or mismatched result never drives a current decision.
RTSP → stamped frames -
02
Follow what moves.
People are detected and followed as tracks across frames, on the Jetson. Empty, fresh, stale and offline states are published explicitly rather than guessed.
Detect + track · on device -
03
Apply the rule.
Zone rules decide what counts as an event. When one does, an incident opens in a local SQLite database, and alerts leave through a restart-safe outbox.
Zone rules → SQLite incident -
04
Review it where it lives.
A dashboard on the device's loopback address shows the live picture, the incidents and their context. You reach it through an SSH tunnel.
127.0.0.1 · SSH tunnel
02 — In practice
Not a screenshot. The dashboard, running.
This is the beta dashboard's own interface, driven by the scripted courtyard above instead of a device. Click through it, ask it something, or steer the scene.
The live dashboard could not be embedded here. Open it in its own tab
00.0 sScene clock
- TryOpen an incident to see what was kept and why.
- TryAsk a question in the Ask panel. Here the answers come from the scripted scene, not from a model.
- TryChoose “Stream stalls” and watch freshness change from fresh to stale.
- TryOpen Features, press “Use the demo key”, and switch a feature off or on.
Synthetic demonstration. No camera, model or device is connected: the picture is an illustration, the people are scripted and the answers are examples. Nothing here measures recognition or model performance.
Evidence review
What an incident keeps.
The moment and the record of it, side by side: the frame as the camera saw it, and the few facts Sentinel writes down about it.
- Event
- Person in restricted zone Zone A · a confirmed track stayed past the rule's threshold
- Track
- 73 one person, followed across frames
- Source time
- 14.0 s of the scripted loop; on a device, the frame's own time
- Stored
- Local SQLite on the device
- Snapshot
- Opt-in kept only when you enable it
- Alert
- Local outbox Telegram is off by default
Freshness is visible.
Fresh, stale, offline and empty states keep the age of an observation in view.
Context is connected.
An incident carries the time and stream stamps of the observation behind it, so a record can be checked against what was seen.
Review stays local.
The dashboard binds to localhost. Nothing about review requires an outside service.
03 — The hardware
1 / 7
Two parts. One local loop.
A camera to observe and an edge computer to process. In the featured build they are an Imou Ranger S2 and a Jetson Orin Nano Super.
2 / 7
It starts at the glass.
One RTSP stream enters the pipeline. The dashboard's preview reuses those captured frames; it does not open a second connection to the camera.
3 / 7
One observation.
A frame arrives with its source time attached, so current activity stays distinct from stale frames.
4 / 7
The same frame, opened.
Lifted out of the camera's view so the scene can be read: the yard, the person, the zone, and the cone the camera sees. An illustration: Sentinel works on the flat frame.
5 / 7
The event stays identifiable.
The person is followed as one track. When that track stays in Zone A past the rule's threshold, the zone rule opens an incident for it.
6 / 7
Carried to the board beside it.
Detection, tracking and incident storage run on the Jetson, on your own network. External alerts are a separate choice, off by default.
7 / 7
Local evidence.
An incident record in local SQLite, stamped with the time of the observation behind it. Snapshots are kept only if you turn them on.
- Event
- Person in restricted zone · Zone A · track 73
- Stored
- Local SQLite, on the device
- Alert
- Local outbox · Telegram off by default
Illustrative hardware study · synthetic courtyard · no camera connectedIllustrative study · synthetic scene · no camera
The eyes
Imou Ranger S2
The camera in the featured build. Its RTSP stream is the starting point for the local pipeline.
The compute
Jetson Orin Nano Super
The edge computer beside the camera. Detection, tracking and incident storage run on the Jetson.
The models in this sequence are original, generic studies of a pan-tilt camera and a compute board, not renders of those products. Featured setup, not a compatibility certification.
04 — Built in the open · v2.0.0-beta.1
A beta, with clear boundaries.
Sentinel is a project you can inspect, test and help improve. Start with the source and the limits of the current beta.
Implemented
The core one-camera path: detection, tracking, zone rules, local SQLite incidents, an alert outbox and a dashboard. Portable tests and a bounded device check exist. The system has no production acceptance.
Experimental
Face recognition and visual chat have not passed admission for normal use. The dashboard may show these panels. No accuracy, speed or reliability guarantee is claimed.
Unverified
The reference is a Jetson Orin Nano 8 GB. A complete launch from the current tester package, and any other Jetson configuration, remain unverified. Read the tester guide and release notes before trying the beta.
What leaves the device
Core inference and incident storage are local. External alert channels are disabled by default; enabling Telegram sends selected alert information to that external service. This website does not connect to your camera.